What RNZ reported

On 11 November 2026, RNZ's Nine to Noon spoke with Saskia de Vilder about a programme that aims to map where fast-moving landslides are likely to hit after a big quake or a heavy storm. De Vilder is an engineering geologist at Earth Sciences New Zealand. The programme is called Hōretireti Whenua, or Sliding Lands. Earth Sciences NZ and Massey University run it together. It uses existing landslide records, machine learning and other AI methods to estimate how likely a slope is to fail under a given amount of rain or shaking.

The need is real. New Zealand doesn't yet have landslide hazard models that apply across the whole country. In June, Chris Massey of Earth Sciences NZ told Engineering New Zealand that, as a result, "we don't really know the full scale of the impact of landslides across the country." January 2026 showed what that gap can cost. Two rain-triggered slips in Tauranga, one at Mount Maunganui and one at Pāpāmoa, killed eight people. The same Engineering New Zealand article notes that over the past two centuries, landslides have killed more people here than earthquakes, eruptions and tsunamis combined.

Who's behind it and how it's funded

Sliding Lands is an MBIE Endeavour Fund programme that runs from 2023 to 2028. It has three leads: de Vilder, Chris Massey and Emma Hudson-Doyle of Massey University. De Vilder and Massey began the work at GNS Science, which is now part of Earth Sciences NZ. Neither the programme page nor RNZ gives a dollar figure, so we can't say how much public money is involved.

How the model fits together

The programme page splits the work into four linked strands:

  • Triggering scenarios: the rainfall and earthquake conditions that start slopes moving.
  • Geospatial susceptibility: which parts of the country are most likely to fail, based on terrain, geology, vegetation and similar factors.
  • Size and runout: how big a slide of each type is likely to be and how far its debris travels.
  • Vulnerability: what a slide does to the people, buildings and infrastructure in its path.

De Vilder told RNZ the team divides regions into "landslide domains." Each domain groups places with similar climate, geology and rainfall. Before a storm, the model can take a rainfall forecast and turn it into a map of where slips are most likely. The programme says its end products will be landslide nowcasts and impact forecasts for decision-makers.

She described uses at three stages. Before an event, the maps could help emergency managers decide where to station crews. During an event, they could show responders which places to watch. Afterwards, they could guide reconnaissance flights and field checks. Speed matters here. In de Vilder's words, the fast landslides this programme targets can leave "limited or no evacuation time."

Where the training data comes from

Cyclone Gabrielle left researchers with a huge dataset. A study led by Chris Massey, published in the journal Landslides in August 2025, counted more than 800,000 slips from that storm. Together they covered about 100 square kilometres across Tairāwhiti and Hawke's Bay. University of Canterbury students, led by Tom Robinson, spent nine months mapping most of them. In scrub and grassland there were up to 327 landslides per square kilometre. Inland from Esk Valley, 500 mm of rain fell in 24 hours, more than 50% above the previous record.

After rainfall, the factors that mattered most were geology, vegetation type and how steep the slope was. Forest held up better than scrub and pasture, and native forest did better than exotic plantations. Land that lost its vegetation three to five years before the cyclone failed more often. Sliding Lands helped fund the study, and Earth Sciences NZ says the data will feed into its national models.

A related model shows the pitfalls

A separate national model gives a useful early look at what can go wrong. In February 2026, Oliver Wigmore posted a preprint to EGUsphere. He works at the Department of Conservation and at Victoria University of Wellington's Antarctic Research Centre. His team found Gabrielle landslides in Sentinel-2 satellite imagery and kept only those that also appeared in a Dragonfly Data Science inventory. They trained gradient-boosted decision trees on those slips and produced susceptibility maps at 25 m resolution for the whole country, under current and future storm scenarios. The work isn't part of Sliding Lands, but it uses similar methods.

The headline score was a ROC-AUC of 0.94. Wigmore reported that under warmer scenarios, the area at risk grows faster than rainfall intensity does. The public referee reports raised serious concerns:

  • One reviewer said random cross-validation ignores spatial autocorrelation, so pixels that sit next to each other can end up in both the training and test sets.
  • Another said precision-recall for the landslide class was fairly weak, and that at a 5% threshold, false positives outnumbered true ones.
  • Both asked whether a model trained on one storm in one region can say anything reliable about the Southern Alps, where the rocks and slope processes are different.

In July, Wigmore replied that he'd added spatial cross-validation and would revise the paper. The revised numbers aren't public yet.

What we still don't know

  • Accuracy: Sliding Lands hasn't published performance figures for its models.
  • Openness: De Vilder said users will be able to "interrogate" the model, seeing where machine learning was used and which past landslides shaped it. We haven't seen a confirmed plan to release the data or code.
  • Resolution: It's a screening tool. It can't tell homeowners whether their own section is safe, and de Vilder said finer detail depends on how the output gets used.
  • Adoption: She called talks with councils and other end users about adding the model to hazard maps "ongoing."
  • Timing: Funding runs to 2028, but there's no public launch date.

What builders can take from this

If you're building geospatial machine learning in New Zealand, this work offers some practical lessons:

  1. Report precision-recall as well as ROC-AUC. Landslides are rare compared with stable ground, so a high ROC-AUC can hide a lot of false alarms.
  2. Validate by place. Use spatial blocks, or hold out whole regions and whole events. A model that has only seen Gabrielle should be tested on a different storm somewhere else.
  3. Record where your labels came from. Satellite change detection picks up any bare ground, which includes slips, runout and other disturbance. Note how each label was made.
  4. Check the physics. SHAP plots and review by domain experts help catch effects that your own filters created, such as a slope cutoff that builds a slope signal into the labels.
  5. Build for the people who'll use it. Councils, lifeline utilities, insurers and forestry companies will need tools that turn national screening maps into decisions about crews, roads and assets. Small teams can do useful work connecting the two.

We'll keep following Sliding Lands as its outputs come out.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member